Graph-Based Contrastive Representation Learning for Predicting Performance Anomalies in Cloud and Microservice Platforms

This paper proposes a self-supervised modeling framework based on contrastive time-series representation learning to address the complexity of backend system performance anomaly prediction in cloud computing and microservice environments. The method constructs a time-varying service dependency graph and a temporal encoding mechanism to achieve joint representation of spatial structural features and temporal dynamic features, enabling the unsupervised identification of potential performance degradation patterns. The model consists of four main components: a dynamic graph construction module, a graph convolution feature extraction module, a time-series encoding module, and a contrastive learning optimization module. The dynamic graph module captures the evolving dependencies among services, while the time-series encoding module extracts multi-scale temporal features. The contrastive learning module builds positive and negative sample pairs to achieve representation aggregation and differentiation in the latent space. Extensive experiments on real backend system monitoring datasets, along with sensitivity analyses on learning rate, optimizer, temperature coefficient, and data missing rate, demonstrate that the proposed model outperforms mainstream methods in accuracy, precision, recall, and AUC, showing strong generalization and robustness. This study provides a new technical approach for early identification of performance anomalies in complex distributed systems and offers practical, theoretical, and methodological support for intelligent operation and performance assurance in cloud platforms.